An adaptive analgesic pump administration control method and system

Abstract

The application discloses a kind of self-adapting analgesia pump administration control method and system, belong to medical administration technical field.Method includes: the real-time physiological parameter sequence of patient and analgesia evaluation data are collected as input data set;Time-frequency transformation is executed to input data set to extract pain-related rhythm fluctuation component and construct multidimensional time-frequency feature tensor;Characteristic gradient field is calculated on multidimensional time-frequency feature tensor and executes manifold learning dimension reduction to extract the key manifold structure of pain event;Key manifold structure is input into deep reinforcement learning model to make policy network output reference administration action sequence;Difference comparison is generated dynamic administration adjustment amount by reference administration action sequence and real-time physiological parameter sequence;Dynamic administration adjustment amount and historical administration record cumulative fusion calculate current instruction administration rate;Instruction administration rate is converted into the driving pulse sequence of step motor and is output to analgesia pump actuator.The application realizes the accurate self-adapting control of analgesia pump administration.

Description

Technical Field

[0001] This invention relates to the field of medical technology, specifically to an adaptive analgesic pump drug delivery control method and system. Background Technology

[0002] Existing analgesic pump drug delivery control methods primarily rely on patient-controlled analgesia (PCA) or continuous infusion based on a fixed dosing regimen. PCA requires the patient to actively press a button when experiencing pain, with the timing and dosage entirely dependent on the patient's subjective feelings and responsiveness. Continuous infusion, on the other hand, delivers analgesics at a preset, fixed rate, unable to dynamically adjust based on real-time changes in the patient's pain status. Both of these traditional methods fail to adequately consider the objective pain information inherent in the patient's physiological parameters, leading to a time lag and dosage deviation between dosing decisions and the patient's actual analgesic needs. Existing technologies have significant drawbacks. Patient-controlled dosing, based on subjective pressure, cannot be actively triggered when the patient is not fully awake from surgical anesthesia, is sedated, or is confused, easily leading to analgesia interruption and sudden pain. Furthermore, different patients exhibit significant differences in their pain tolerance thresholds and expression methods, making it difficult to objectively quantify the true intensity and changing trends of analgesic needs based on subjective ratings. Continuous infusion is completely incompatible with the dynamic physiological processes of patients, such as fluctuations in blood drug concentrations and changes in pain rhythms. Fixed-rate dosing may lead to serious risks such as insufficient analgesia causing pain or excessive analgesia causing respiratory depression. Current technologies lack effective means to extract and utilize the pain characteristic rhythms contained in the patient's continuous physiological signals, making it impossible to accurately identify the precursors and evolution patterns of pain events. At the same time, existing dosing decision-making mechanisms are insufficient for flexible and rapid adjustments based on real-time monitoring of physiological status feedback, and the precision of dosing rate and dosage control is inadequate.

[0003] This invention addresses two key issues. First, existing technologies struggle to accurately extract pain-related dynamic rhythmic features from multimodal physiological parameters, failing to effectively identify objective physiological signal patterns preceding pain events. Second, current dosing decision-making methods lack adaptive dynamic adjustment mechanisms based on real-time physiological feedback, resulting in lag in dosing rate adjustments and insufficient optimization capabilities. Summary of the Invention

[0004] This invention aims to address the technical problem of inaccurate and untimely extraction of dynamic pain features implicit in real-time physiological parameters in existing analgesic pump drug delivery control methods. By performing time-frequency transformation on the collected multimodal physiological parameters, specific frequency band components related to pain rhythm are screened out, and a multidimensional time-frequency feature tensor is constructed to achieve accurate mapping and structured expression of pain features from the time domain to the time-frequency domain. This invention also aims to address the technical problem of existing drug delivery decision-making methods lacking dynamic adaptive capabilities based on objective physiological feedback. By inputting the extracted key manifold structure into a deep reinforcement learning model to generate a baseline action sequence, and further comparing it with the real-time physiological state vector through differential comparison to generate dynamic adjustment quantities, the final commanded drug delivery rate is calculated by fusing historical drug delivery records, thus constructing a closed-loop feedback mechanism from pain feature identification to drug delivery rate optimization.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This invention provides an adaptive analgesic pump drug delivery control method and system, aiming to solve the problems of delayed or excessive drug delivery caused by the lack of real-time and accurate perception of the patient's pain status in existing analgesic pump drug delivery methods. This method achieves dynamic adaptive adjustment of the drug delivery rate through the fusion analysis of multimodal physiological signals and subjective assessment data, combined with deep learning and manifold learning techniques, significantly improving analgesic efficacy and drug safety.

[0007] Specifically, real-time physiological parameter sequences and analgesia assessment data from patients are collected as input datasets. Heart rate, blood pressure, skin conductance, and blood oxygen saturation signals are simultaneously recorded within continuous time sampling windows and stitched together in time alignment to form a real-time physiological parameter sequence. Simultaneously, at the end of each sampling window, a visual simulation analgesia score triggered by the patient is acquired. This score is then bound to the corresponding window's physiological parameter sequence and organized into a two-dimensional data matrix. This process integrates objective physiological indicators with subjective pain perception, providing a comprehensive and timely data foundation for subsequent feature extraction.

[0008] A time-frequency transform was performed on the input dataset to extract pain-related rhythmic fluctuation components and construct a multidimensional time-frequency feature tensor. Heart rate and skin conductance signals were extracted from the two-dimensional data matrix to form time-series curves. Short-time Fourier transforms were performed on both to obtain time-spectrum matrices. Rhythmic components with a frequency range of 0.05Hz to 0.15Hz were selected from the heart rate time-spectrum matrix as heart rate pain rhythm components, and rhythmic components with a frequency range of 0.1Hz to 0.5Hz were selected from the skin conductance time-spectrum matrix as conductance pain rhythm components. These two components were then stacked along the frequency dimension to construct a three-dimensional time-frequency feature tensor. The first dimension of this tensor is the time window index, the second dimension is the frequency point index, and the third dimension is the physiological parameter type index, thus effectively preserving the pain-related rhythmic features and their time-varying patterns, providing a structured input for accurately characterizing pain state transitions.

[0009] A feature gradient field is computed on a multidimensional time-frequency feature tensor, and manifold learning is performed for dimensionality reduction to extract the key manifold structure of pain events. Feature difference vectors between adjacent windows are computed along the time window dimension and combined to form a feature gradient field reflecting the dynamic changing trend of pain features. A local linear embedding algorithm is performed on the feature gradient field to project the high-dimensional feature space onto a two-dimensional low-dimensional embedding space, thereby eliminating redundant noise and revealing the intrinsic manifold structure. In the two-dimensional embedding space, the local density value of each embedding point is computed: for the... Embedding points Select its nearest The neighborhood set consists of 1 neighboring points. According to the formula:

[0010]

[0011] in: For the number of neighboring points, For Euclidean distance, For adaptive bandwidth parameters (equal to) To its first (Distance between nearest neighbors). Time windows corresponding to embedding points whose local density values ​​exceed a preset threshold are marked as candidate windows for pain events. The corresponding feature vectors of these windows in the three-dimensional time-frequency feature tensor are extracted and organized into a low-dimensional manifold structure as the key manifold structure for pain events. This manifold structure can effectively identify typical spatiotemporal patterns of pain attacks, avoid the sensitivity of traditional thresholding methods to individual differences, and improve the robustness of pain event detection.

[0012] The key manifold structure is input into a deep reinforcement learning model, and a baseline drug administration sequence is output through a policy network. Specifically, each feature vector in the key manifold structure is sequentially input into a state encoder to be mapped to a hidden state vector, and then input into the policy network to output the Gaussian distribution of the action mean and variance parameters. Normalized drug administration action values ​​are sampled from this distribution, and multiplied by a preset maximum drug administration rate to obtain the instantaneous drug administration rate corresponding to the current state. The instantaneous drug administration rates of all windows are arranged in chronological order to form the baseline drug administration sequence. The deep reinforcement learning model can learn the optimal drug administration strategy in complex pain dynamic environments, and the output baseline sequence initially meets the general analgesic needs of patients.

[0013] A dynamic dosing adjustment is generated by differentially comparing a baseline dosing sequence with a real-time physiological parameter sequence. A physiological state vector is formed by extracting heart rate, blood pressure, and skin conductance values ​​from three consecutive sampling points prior to the current time window from the real-time physiological parameter sequence. A baseline dosing rate is extracted from the baseline action sequence. The L2 norm of the vector difference between the physiological state vector and the pre-stored ideal physiological state vector is calculated and multiplied by a preset proportional gain coefficient to obtain the first adjustment component. Simultaneously, the difference between the baseline dosing rate and the actual dosing rate in the previous time window is calculated and multiplied by a preset integral gain coefficient to obtain the second adjustment component. The two are then added together to obtain the dynamic dosing adjustment. This differential mechanism can correct the deviation between the baseline sequence and the actual physiological state in real time, eliminating the influence of model prediction errors and individual differences.

[0014] The dynamic dosing adjustment is accumulated and fused with historical dosing records to calculate the commanded dosing rate at the current moment. The specific steps are as follows: The cumulative total dosing and real-time dosing rate at the end of the previous time window are read from the historical records; the dynamic adjustment is superimposed on the real-time dosing rate to obtain the dosing rate to be corrected, and a pre-set upper and lower limit is used for limiting; the limited dosing rate is multiplied by the length of the current time window to obtain the planned dosing, and this is added to the cumulative total dosing to obtain the updated cumulative total dosing; it is then determined whether the updated cumulative total dosing exceeds the safe total dosing threshold: if it does not, the limited dosing rate is directly used as the commanded dosing rate; if it does, the commanded dosing rate is recalculated according to the safe total dosing threshold so that the planned dosing exactly fills the remaining capacity. During recalculation, the cumulative total dosing before the start of the current time window is set as follows: The safe total dosage threshold is The window duration is The elapsed time is Then the remaining capacity Remaining time The recalculated command dosing rate The unit is milliliters per hour. This safety limit and cumulative constraint mechanism ensures that the total dosage is always below the safety limit, effectively avoiding the risks caused by drug overdose.

[0015] The commanded dosing rate is converted into a stepper motor drive pulse sequence and output to the actuator of the analgesia pump. The volume of medication administered and the number of drive pulses corresponding to one revolution of the stepper motor are obtained. The required drive pulse frequency per second is calculated based on the commanded dosing rate, converted into the time interval between adjacent pulses, and a square wave pulse sequence with a fixed pulse width is generated and output to the driver input port. The stepper motor rotates by a corresponding angle according to the number of pulses, pushing the analgesia pump piston to move the corresponding distance, achieving precise micro-dose administration. This conversion process directly maps digital commands to mechanical motion, ensuring the response speed and accuracy of drug administration.

[0016] The present invention also includes an adaptive analgesic pump drug delivery control system, which includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned adaptive analgesic pump drug delivery control method. This system integrates all the above-mentioned technical aspects, enabling automated and intelligent analgesic pump drug delivery control in a real clinical environment.

[0017] The beneficial effects of this invention are:

[0018] A time-frequency transform is performed on the input dataset to extract pain-related rhythmic fluctuation components and construct a multidimensional time-frequency feature tensor. By mapping the heart rate and skin conductance time-series curves to the time-frequency domain using short-time Fourier transform, characteristic frequency band components related to sympathetic nerve excitation—namely, the heart rate pain rhythm component (0.05 Hz to 0.15 Hz) and the conductance pain rhythm component (0.1 Hz to 0.5 Hz)—can be separated from complex, non-stationary physiological signals. This method overcomes the limitation of conventional time-domain analysis methods in capturing subtle rhythmic changes caused by pain, enabling the precise quantification and localization of pain-related physiological fluctuation features in the frequency dimension. The selected rhythmic components are stacked in the frequency dimension to construct a three-dimensional time-frequency feature tensor, preserving the complete information structure of time, frequency, and physiological parameter type. This provides a richer and clearer data organization for subsequent feature analysis, significantly improving the accuracy and robustness of pain feature extraction. The feature gradient field is calculated on the multidimensional time-frequency feature tensor, and manifold learning is performed for dimensionality reduction to extract the key manifold structure of pain events. By constructing a feature gradient field through the feature difference vectors between adjacent time windows, the changing trends and fluctuations of physiological signals over time can be intuitively depicted, highlighting the rapid transitions in the feature space before and after the occurrence of pain events. A local linear embedding algorithm is employed to project the high-dimensional feature gradient field onto a two-dimensional low-dimensional embedding space, effectively simplifying and visualizing the inherent topological structure of the complex high-dimensional pain feature space. In the low-dimensional embedding space, by calculating the local density value of each embedding point and setting an adaptive bandwidth parameter, typical pattern regions representing the occurrence of pain events—i.e., high-density clusters—can be accurately identified based on the local density of the data distribution. The feature vectors of the time windows corresponding to these high-density clusters are extracted as key manifold structures, achieving efficient screening and noise reduction of key features of pain events and eliminating interference from physiological signals in non-painful states. This processing results in lower-dimensional, more semantically explicit, and less noisy pain feature information input into subsequent deep reinforcement learning models, effectively reducing the complexity of model processing and improving the accuracy of state representation.

[0019] A dynamic dosing adjustment is generated by differentially comparing the baseline dosing sequence with the real-time physiological parameter sequence. A physiological state vector is constructed by extracting heart rate, blood pressure, and skin conductance values ​​from three consecutive sampling points prior to the current time window from the real-time physiological parameter sequence. This vector captures rapid changes in the patient's recent physiological state, providing immediate physiological feedback. The L2 norm of the vector difference between this physiological state vector and the pre-stored ideal physiological state vector is calculated and multiplied by a proportional gain coefficient to obtain the first adjustment component. This directly quantifies the deviation of the current physiological state from the ideal steady state, making the adjustment directly correlated with the severity of physiological imbalance. Simultaneously, the difference between the baseline dosing rate and the actual dosing rate in the previous time window is calculated and multiplied by an integral gain coefficient to obtain the second adjustment component. This introduces historical memory of the dosing rate change trend, effectively suppressing system oscillations caused by rate abrupt changes. The proportional and integral adjustment components are added to generate a comprehensive dynamic dosing adjustment, achieving real-time, smooth, and gradual correction of the baseline dosing rate. This ensures that the final dosing rate tracks physiological state changes while maintaining system stability.

[0020] The commanded dosing rate is converted into a sequence of drive pulses for a stepper motor and output to the actuator of the analgesic pump. By calculating the volume of medication and the number of drive pulses corresponding to each rotation of the stepper motor, the continuous commanded dosing rate is discretized into a specific drive pulse frequency. The pulse frequency is then converted into a precise time interval between adjacent pulses, and a square wave pulse sequence with a fixed pulse width is generated according to this time interval. This method of converting digital commands into physical drive pulses achieves precise open-loop control of the motor's rotation angle and speed without the need for additional position feedback devices. The stepper motor rotates the corresponding angle strictly according to the number of pulses in the pulse sequence, directly pushing the analgesic pump piston to move the corresponding linear distance, thereby accurately converting the electrical pulse signal into linear displacement and medication volume. This method ensures high repeatability and controllability of the medication volume, avoids dosage inaccuracies caused by motor control errors, and improves the reliability and accuracy of dosing execution. Attached Figure Description

[0021] The invention will now be further described with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart of an adaptive analgesic pump drug delivery control method according to the present invention;

[0023] Figure 2 This is a flowchart of the construction of pain-related multidimensional time-frequency feature tensors;

[0024] Figure 3 This is a flowchart of key manifold structure extraction for pain events based on manifold learning. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] See Figure 1 This invention provides an adaptive analgesic pump drug delivery control method, comprising: collecting a patient's real-time physiological parameter sequence and analgesia assessment data as an input dataset; performing time-frequency transformation on the input dataset to extract pain-related rhythmic fluctuation components and constructing a multidimensional time-frequency feature tensor; calculating a feature gradient field on the multidimensional time-frequency feature tensor and performing manifold learning dimensionality reduction to extract key manifold structures of pain events; inputting the key manifold structures into a deep reinforcement learning model and causing the policy network of the deep reinforcement learning model to output a baseline drug delivery action sequence; performing differential comparison between the baseline drug delivery action sequence and the real-time physiological parameter sequence to generate a dynamic drug delivery adjustment amount; accumulating and fusing the dynamic drug delivery adjustment amount with historical drug delivery records to calculate the commanded drug delivery rate at the current moment; and converting the commanded drug delivery rate into a stepper motor drive pulse sequence and outputting it to the actuator of the analgesic pump.

[0027] Example 1: In specific implementation, when collecting the patient's real-time physiological parameter sequence and analgesia assessment data as the input dataset, the patient's heart rate, blood pressure, skin conductance, and blood oxygen saturation signals are simultaneously recorded within a continuous time sampling window. The heart rate, blood pressure, skin conductance, and blood oxygen saturation signals are then concatenated into a real-time physiological parameter sequence after being aligned by time. At the end of each sampling window, an analgesia visual analog scale (VAS) score triggered by the patient is obtained. The VAS score is then bound to the corresponding real-time physiological parameter sequence of the sampling window. The bound real-time physiological parameter sequence and VAS score are organized into a two-dimensional data matrix. This two-dimensional data matrix is ​​used as the input dataset.

[0028] In some embodiments, the duration of the continuous-time sampling window is set to a fixed value, and four physiological parameter signals are simultaneously acquired within each sampling window. The heart rate signal is acquired through ECG electrodes or photoplethysmography (PPG) sensors; the blood pressure signal is acquired through a non-invasive blood pressure cuff or continuous blood pressure monitoring device; the skin conductance signal is acquired through electrodes placed on the palm or forearm; and the blood oxygen saturation signal is acquired through a pulse oximeter probe. The reason why the blood oxygen saturation signal is included as one of the four mandatory signals during the acquisition phase is that blood oxygen saturation is a fundamental indicator for monitoring the vital signs of postoperative patients. Continuous monitoring of blood oxygen saturation during analgesia pump administration can promptly detect the risk of respiratory depression that may be caused by excessive analgesia, providing the system with a safety early warning capability. All signals are timestamped by the same clock source during acquisition for subsequent time alignment. Time alignment means that the four signal values ​​at each sampling moment are matched one-to-one according to the timestamp. If a signal has a missing value, linear interpolation is used to fill in the missing data. The four aligned signals are arranged in chronological order. The four values ​​at each sampling time form a four-dimensional vector, and the four-dimensional vectors at all sampling times constitute the real-time physiological parameter sequence in chronological order.

[0029] In some embodiments, the visual analog scale (VAS) score for analgesia is input by the patient on an electronic pain scoring terminal via swiping or clicking. The score ranges from 0 to 10, where 0 represents no pain and 10 represents the most severe pain. At the end of each sampling window, the system prompts the patient to input the pain score for that window. After the patient completes the input, the score is recorded and bound to the real-time physiological parameter sequence of the current sampling window. The binding operation involves appending the score values ​​from the same sampling window as an additional data column to the end of the real-time physiological parameter sequence, expanding the vector at each sampling moment into a five-dimensional vector (containing four physiological parameter values ​​and one score value). All five-dimensional vectors are organized chronologically into a two-dimensional data matrix, where rows correspond to different sampling moments and columns correspond to the five data dimensions (heart rate, blood pressure, skin conductance, blood oxygen saturation, and VAS score for analgesia). This two-dimensional data matrix constitutes the input dataset.

[0030] Example 2: In specific implementation, refer to Figure 2 When performing time-frequency transformation on the input dataset to extract pain-related rhythmic fluctuation components and construct a multidimensional time-frequency feature tensor, the heart rate signal is extracted from the two-dimensional data matrix to form a heart rate time-series curve. The skin conductance signal is extracted from the two-dimensional data matrix to form a skin conductance time-series curve. Short-time Fourier transforms are performed on the heart rate time-series curve and the skin conductance time-series curve, respectively, to obtain the heart rate time-spectrum matrix and the skin conductance time-spectrum matrix.

[0031] In some embodiments, the short-time Fourier transform uses a Hanning window as the window function, with a window function length of 256 sampling points and a window overlap rate of 50%. The rows of the heart rate time-spectrum matrix correspond to time window indices, the columns correspond to frequency point indices, and the matrix element values ​​are the amplitude spectrum values ​​at the corresponding time window and frequency point. The skin conductance time-spectrum matrix has the same row and column structure.

[0032] From the heart rate frequency spectrum matrix, rhythmic fluctuation components with a frequency band ranging from 0.05 Hz to 0.15 Hz are selected as the heart rate pain rhythm components. The selection method is as follows: all columns in the heart rate frequency spectrum matrix whose frequency values ​​corresponding to the frequency indices are located within the interval [0.05 Hz, 0.15 Hz] are retained, resulting in a new submatrix. The number of rows in this submatrix remains unchanged, and the number of columns is equal to the number of frequency points contained within the frequency band. This submatrix is ​​the heart rate pain rhythm component.

[0033] From the skin conductance time-spectrum matrix, rhythmic fluctuation components with a frequency band ranging from 0.1 Hz to 0.5 Hz are selected as the conductance pain rhythm components. The selection method is as follows: all columns in the skin conductance time-spectrum matrix whose frequency values ​​corresponding to the frequency indices are located within the interval [0.1 Hz, 0.5 Hz] are retained, resulting in a new sub-matrix. The number of rows in this sub-matrix remains unchanged, and the number of columns is equal to the number of frequency points contained in the frequency band. This sub-matrix is ​​the conductance pain rhythm component.

[0034] The heart rate pain rhythm component and the electrical conductance pain rhythm component are stacked in the frequency dimension to construct a three-dimensional time-frequency feature tensor. The stacking operation refers to treating the two sub-matrices as two two-dimensional matrices, each with dimensions of (number of time windows, number of frequency points), and merging them in the third dimension to form a three-dimensional array. The first dimension of the three-dimensional time-frequency feature tensor is the time window index, corresponding to the time window number of the short-time Fourier transform; the second dimension is the frequency point index, corresponding to the discrete frequency point number within the filtered frequency band; the third dimension is the physiological parameter type index, where index value 1 corresponds to the heart rate pain rhythm component, and index value 2 corresponds to the electrical conductance pain rhythm component. Each element of the three-dimensional time-frequency feature tensor stores the amplitude spectrum value corresponding to the time window, the frequency point, and the physiological parameter type.

[0035] Example 3: In a specific implementation, when calculating the feature gradient field on the multidimensional time-frequency feature tensor and performing manifold learning dimensionality reduction to extract the key manifold structure of the pain event, refer to... Figure 3 The feature difference vector between adjacent time windows is calculated along the time window dimension of the three-dimensional time-frequency feature tensor. The first dimension of the three-dimensional time-frequency feature tensor is the time window index, and the total number of time windows is denoted as . For the first A feature matrix is ​​extracted from the three-dimensional time-frequency feature tensor within a time window. The dimension of this matrix is ,in The number of frequency point indices, Number of physiological parameter type indices (fixed to 2). Feature difference vector between adjacent time windows. Through calculation and The difference between each corresponding element is obtained by flattening all differences into a one-dimensional vector, where From 2 to . The dimension is .

[0036] The feature difference vectors from all time windows are combined into a single feature gradient field. The combination method is as follows: Feature difference vectors Arranged in chronological order, forming a matrix. Its number of rows is The number of columns is The matrix This is the characteristic gradient field.

[0037] A local linear embedding algorithm is performed on the feature gradient field to project the high-dimensional feature space onto a two-dimensional low-dimensional embedding space. The input to the local linear embedding algorithm is the feature gradient field matrix. Each row vector represents a sample point (corresponding to the gradient within a time window), and the total number of samples is... The dimension of each sample is Local linear embedding algorithm sets neighborhood parameters. The value is 12. The reconstruction weights are calculated by minimizing the local reconstruction error, and the output is the coordinates of each sample point in the two-dimensional low-dimensional embedding space, resulting in a... matrix , of which The row corresponds to the first The feature difference vector (i.e. the first feature difference vector) Embedding points of the gradient between each time window and the previous time window. ,index From 1 to The core premise of the locally linear embedding algorithm is that high-dimensional data is distributed on a sufficiently smooth low-dimensional manifold, and neighboring points lie within a locally linear slice on the manifold. In this invention, the constructed feature gradient field originates from continuously sampled temporal signals of physiological parameters. The feature difference vectors between adjacent time windows form a continuous trajectory in the time dimension. Because physiological parameters are constrained by neurohumoral regulation mechanisms, their changes exhibit physiological continuity and do not abruptly change. Therefore, the feature gradient vectors of adjacent time windows are adjacent to each other in the feature space, satisfying the requirement of smoothness in the local neighborhood for the locally linear embedding algorithm. The occurrence, development, and resolution of a pain event is a continuous physiological process. Although the feature gradient vectors at different stages may differ in direction and amplitude, these difference vectors form a continuous manifold trajectory in the feature space. The locally linear embedding algorithm utilizes this continuity of the local neighborhood to project the high-dimensional feature gradient field onto a low-dimensional embedding space and identifies the manifold region corresponding to the pain event through local density filtering.

[0038] Calculate the local density value for each embedding point in the two-dimensional low-dimensional embedding space. For the first embedding point in the two-dimensional low-dimensional embedding space... Embedding points Select its nearest The neighborhood set consists of 1 neighboring points. .in Set to 15. The value is determined based on the total number of embedding points. When it is greater than 100, take A value of 15 allows for stable estimation of local density; if If less than 100, then Take as ,but The minimum value is not less than 5. Calculate the number according to the following formula. Local density value of each embedding point :

[0039]

[0040] in: For neighborhood set The number of points in the middle is equal to ; For embedding point and The Euclidean distance between them; For the first The adaptive bandwidth parameter for each embedding point is equal to... To its first Distance between nearest neighbors Indicates to Round down when hour, ,Right now equal The Euclidean distance to its 7th nearest neighbor.

[0041] The time windows corresponding to embedding points where the local density value exceeds a preset density threshold are marked as candidate windows for pain events. The preset density threshold is set to all local density values. ( It is 1.2 times the average of all (). The specific calculation method is: first calculate all mean Then set the threshold. For each embedding point ,like Then the feature difference vector corresponding to the embedding point Corresponding time window index Marked as a candidate window for pain events. Note the feature difference vector. The corresponding time window is the first One, because By the The and the first The feature matrix of the nth time window is calculated. Simultaneously, the nth... The window preceding the first time window (i.e., the first...) Each time window is also marked as a candidate window to avoid missing starting points. Therefore, the set of marked pain event candidate windows contains all those that meet the criteria. of and After deduplication, a series of time window indices are obtained.

[0042] Extract the corresponding feature vectors of all pain event candidate windows into the three-dimensional time-frequency feature tensor. For each time window index labeled as a candidate window... Extract the corresponding feature matrix from the three-dimensional time-frequency feature tensor. (dimension) ), and flatten the matrix into a one-dimensional eigenvector with a length of . Collect the feature vectors of all candidate windows and arrange them in chronological order.

[0043] All extracted feature vectors are organized into a low-dimensional manifold structure as the key manifold structure for the pain event. The organization method is as follows: all feature vectors are treated as row vectors, forming a matrix with the number of rows equal to the number of candidate windows and the number of columns equal to the number of columns. This matrix represents the key manifold structure. This key manifold structure preserves the information of the pain event candidate window in the original time-frequency feature space, while filtering out non-pain-related time periods through manifold learning dimensionality reduction and density filtering.

[0044] Example 4: In a specific implementation, when the key manifold structure is input into a deep reinforcement learning model and the policy network of the deep reinforcement learning model outputs a baseline drug administration sequence, each feature vector in the key manifold structure is sequentially input into the state encoder of the deep reinforcement learning model. The key manifold structure contains multiple feature vectors, and each feature vector corresponds to a pain event candidate window.

[0045] The deep reinforcement learning model employs a deep deterministic policy gradient model based on an actor-critic architecture. Training data is derived from patient physiological parameter sequences and corresponding analgesia assessment data collected from historical clinical records. Each training sample represents a complete treatment cycle, including all real-time physiological parameter sequences from the start to the discontinuation of analgesia pump administration, analgesia visual analog scale (VAS) score sequences, and records of the administration rate manually adjusted by the clinician based on the patient's actual condition. The training samples are labeled as follows: the clinician-adjusted administration rate is used as the optimal action label for each time window, and the patient's VAS score at the end of that time window is used as the evaluation criterion for the action's effectiveness.

[0046] At each time step, after the deep reinforcement learning model outputs a drug administration action, the system calculates a reward value based on the changes in the patient's visual analog scale (VAS) score for analgesia and the changes in the physiological state vector within the next time window following the execution of that action. The reward value consists of two parts. The first part is the analgesic effect reward, calculated by comparing the VAS score for analgesia in the current time window with that in the previous time window; a positive reward is given if the score decreases, and a negative reward is given if the score increases. The second part is the physiological stability reward, calculated by the L2 norm between the physiological state vector in the current time window and the ideal physiological state vector; a positive reward is given if the L2 norm decreases, and a negative reward is given if the L2 norm increases. The total reward value is a weighted sum of the analgesic effect reward and the physiological stability reward, with the weight coefficient of the analgesic effect reward being greater than that of the physiological stability reward.

[0047] A batch of historical treatment cycle data was randomly selected from the training dataset, with a batch size of 64. The physiological parameter sequences and analgesia assessment data from the selected treatment cycle data were processed using the aforementioned data acquisition method, time-frequency transformation to construct a multidimensional time-frequency feature tensor, calculation of the feature gradient field, and manifold learning dimensionality reduction to extract the key manifold structure of the pain event. This key manifold structure served as the input state sequence for the deep reinforcement learning model. The state sequence was sequentially input into the actor network of the deep reinforcement learning model, and the actor network output the drug administration action corresponding to each state. The drug administration action output by the actor network was then input into the critic network, which output an action value estimate based on the current state and the drug administration action. The reward value for each time step was calculated according to the reward function, and the parameters of the critic network were updated using a temporal difference learning method, gradually approximating the true cumulative reward expectation value of the critic network's action value estimate. The parameters of the actor network were updated using a deterministic policy gradient method, ensuring that the drug administration action output by the actor network obtained a higher action value estimate under the evaluation of the critic network. Repeat the above steps until the average error between the judge's network's action value estimate and the actual cumulative reward is less than a preset convergence threshold. During the training of the deep reinforcement learning model, a behavior regularization mechanism is also introduced to constrain the exploration space of the policy network. Specifically, for each historical treatment cycle data extracted from the training dataset, the key manifold structure is extracted as a state sequence according to the aforementioned method, and the dosing rate manually adjusted by the clinician within the same time window is used as the expert action label. At each policy network parameter update, the mean squared error between the dosing action output by the policy network and the expert action label is simultaneously calculated as the behavior bias loss term.

[0048]

[0049] Where M is the number of samples in the current batch. For the first The state feature vector of each sample For the policy network with current parameters The output of the drug delivery action, These are the expert action tags for clinicians in the same state.

[0050] The behavioral bias loss term is weighted and combined with the policy gradient loss term provided by the critic network to form the comprehensive loss function of the policy network:

[0051]

[0052] in: For the deterministic policy gradient loss based on critic network evaluation, The weight coefficient for the behavior regularization term ranges from 0.01 to 0.1. This behavior regularization term guides the policy network to explore within the safe and effective region known to experts, preventing the network from outputting dosing actions that deviate from clinical safety limits during the initial training phase. Furthermore, before the reinforcement learning fine-tuning phase begins, the policy network is pre-trained using expert action labels from the training dataset. The pre-training phase uses only the behavior bias loss term. As the optimization objective, the mini-batch gradient descent method is used to update the network parameters, with a batch size of 64 and a learning rate of 1×10⁻⁶. −3 The training rounds are set to 50. The learning rate is set to 1×10⁻⁶. −3 The rationale is as follows: This magnitude is a commonly used initial step size in supervised training of gradient-based deep neural networks. During the update of network parameters, it ensures that the pre-training loss function decreases steadily along the steepest descent direction. It avoids loss oscillations due to excessively large step sizes, which could deviate from the expert action distribution, and also prevents insufficient convergence within 50 training cycles due to excessively small step sizes. After pre-training, the parameters of the policy network are used as the initial parameters for the fine-tuning phase. Fine-tuning is then performed according to the aforementioned reinforcement learning training method, bringing the initial output distribution of the policy network closer to the clinically optimal decision-making region, thus accelerating the convergence speed of the subsequent reinforcement learning fine-tuning phase.

[0053] The actor network has the following structure: the input layer receives feature vectors from the key manifold structure; the first hidden layer contains 128 neurons using a modified linear unit activation function (MRU); the second hidden layer contains 64 neurons using a modified MRU; and the output layer contains one neuron using a hyperbolic tangent activation function to constrain the output value between -1 and 1. The critic network has the following structure: the input layer receives feature vectors from the key manifold structure and the drug delivery action output by the actor network; the first hidden layer contains 128 neurons using a modified MRU; the second hidden layer contains 64 neurons using a modified MRU; and the output layer contains one neuron using a linear activation function to output the action value estimate.

[0054] The state encoder consists of a three-layer fully connected neural network, with the number of neurons in the input layer equal to the length of the feature vector (i.e., ...). The first hidden layer has 128 neurons, the second hidden layer has 64 neurons, and the output layer has 32 neurons. Each hidden layer is followed by a modified linear unit activation function. The state encoder maps each feature vector to a hidden state vector, with a dimension of 32. The hidden state vectors are input into the policy network of the deep reinforcement learning model. The policy network consists of two fully connected neural networks: an input layer with 32 neurons, a first hidden layer with 64 neurons, and a second hidden layer with 2 neurons. Each hidden layer is followed by a modified linear unit activation function. The policy network outputs a Gaussian-distributed action mean parameter and action variance parameter, where the action mean parameter is a scalar and the action variance parameter is a positive scalar.

[0055] A normalized drug administration action value is obtained by sampling from the Gaussian distribution. The sampling method is as follows: a random number is generated using a standard normal distribution random number generator. According to the formula Calculate normalized drug administration action values ,in The mean parameter of the action. For the action variance parameter, It follows a standard normal distribution with a mean of 0 and a standard deviation of 1. The range of values ​​is within to If the value exceeds this range, it will be truncated.

[0056] The normalized dosing action value is multiplied by a preset maximum dosing rate to obtain the instantaneous dosing rate corresponding to the current state. The maximum dosing rate is set to 20 ml / hour, a value based on the common safety upper limit for clinical analgesic pumps. Instantaneous dosing rate milliliters per hour.

[0057] The instantaneous dosing rates corresponding to all time windows are arranged in chronological order to form the baseline dosing action sequence. The feature vectors in the key manifold structure correspond sequentially to the time order of the pain event candidate windows. Each feature vector generates an instantaneous dosing rate. These instantaneous dosing rates are arranged in the order of their corresponding time window indices to obtain a sequence, the length of which is equal to the number of candidate windows.

[0058] In specific implementation, when the baseline drug administration sequence is differentially compared with the real-time physiological parameter sequence to generate a dynamic drug administration adjustment dose, the heart rate, blood pressure, and skin conductance values ​​of three consecutive sampling points before the current time window are extracted from the real-time physiological parameter sequence. The real-time physiological parameter sequence includes the heart rate, blood pressure, skin conductance, and blood oxygen saturation values ​​at each sampling moment. The current time window refers to the time window corresponding to the candidate window of the pain event being processed. Three consecutive sampling points before the start time of this window are extracted, and each sampling point is a four-dimensional vector. Since the variation of blood oxygen saturation within the normal analgesic dose range is extremely small, its immediate response sensitivity to pain stimuli is lower than that of heart rate, blood pressure, and skin conductance. The latency of heart rate response to sympathetic nerve excitation is usually within a few seconds, the reflexive increase of blood pressure to pain stimuli can be captured within tens of seconds, and skin conductance directly reflects the sympathetic nerve innervation changes of sweat gland activity. These three indicators together constitute a sensitive observation combination for rapid changes in pain status. A decrease in blood oxygen saturation typically occurs after respiratory depression has already developed, making it a lagging indicator rather than a precursor. Including it in the physiological state vector used to calculate dynamic drug administration adjustments not only fails to provide an early signal of pain changes but also dilutes the differential weights of the other three indicators due to its stable nature under normal conditions, reducing the sensitivity of the physiological state vector to pain state changes. Therefore, heart rate, blood pressure, and skin conductance values ​​are extracted from each sampling point, while blood oxygen saturation is ignored. The heart rate, blood pressure, and skin conductance values ​​from three consecutive sampling points are combined to form a physiological state vector. The combination method is as follows: the three values ​​(heart rate, blood pressure, skin conductance) from the first sampling point are arranged sequentially, followed by the three values ​​from the second sampling point, and then the three values ​​from the third sampling point, forming a 9-dimensional vector.

[0059] The baseline dosing rate corresponding to the current time window is extracted from the baseline dosing action sequence. Each element in the baseline dosing action sequence corresponds one-to-one with a pain event candidate window; the corresponding baseline dosing rate is obtained based on the index of the current time window. .

[0060] Calculate the L2 norm of the vector difference between the stated physiological state vector and the pre-stored ideal physiological state vector. The ideal physiological state vector is a 9-dimensional vector, with each dimension value set clinically: heart rate 72 beats per minute, blood pressure 120 mmHg, and skin conductance 2 microsiemens; the values ​​at the three sampling points are identical. Vector difference. ,in This is the current physiological state vector. The vector represents the ideal physiological state. (2-norm) .

[0061] The first adjustment component is obtained by multiplying the L2 norm by a preset proportional gain coefficient. (Proportional gain coefficient) The adjustment was set to 0.5 ml / hour per unit L2 norm, a coefficient determined through simulation experiments, to ensure that the adjustment amount is appropriately corrected when physiological parameters deviate from ideal values. First adjustment component. .

[0062] Calculate the rate difference between the baseline dosing rate and the actual dosing rate in the previous time window. The actual dosing rate in the previous time window refers to the actual dosing rate executed at the end of that window (i.e., the final commanded dosing rate output to the stepper motor). Rate difference ,in This represents the actual dosing rate in the previous time window.

[0063] The second adjustment component is obtained by multiplying the rate difference by a preset integral gain coefficient. Integral gain coefficient The rate was set to 0.2 ml / h (ml / h), a coefficient used to suppress rate fluctuations, determined through simulation parameter tuning. Second adjustment component. .

[0064] The first adjustment component and the second adjustment component are added together to obtain the dynamic dosing adjustment amount. .

[0065] Example 5: In a specific implementation, when calculating the commanded dosing rate at the current moment by accumulating and fusing the dynamic dosing adjustment amount with historical dosing records, the cumulative dosing amount at the end of the previous time window is read from the historical dosing records. The historical dosing records are stored in system memory and include the cumulative dosing amount and real-time dosing rate at the end of each time window. The real-time dosing rate refers to the dosing rate actually executed and output at the end of the previous time window.

[0066] The real-time dosing rate at the end of the previous time window is read from historical dosing records. The dynamic dosing adjustment is then added to the real-time dosing rate to obtain the dosing rate to be corrected. ,in This is the real-time dosing rate at the end of the previous time window. The amount of medication to be adjusted for dynamic dosing.

[0067] The dosing rate to be corrected is limited using preset upper and lower limits to obtain the limited dosing rate. The upper limit is set to 20 ml / h, and the lower limit is set to 0 ml / h. The limiting rule is: if... The dosing rate after limiting ;like ,but ;otherwise .

[0068] Multiply the dose-limited rate by the length of the current time window to obtain the planned dose for the current time window. The length of the current time window. This is a fixed value, set at 0.25 hours (i.e., 15 minutes). Planned dosage. .

[0069] The planned dosage is added to the cumulative dosage to obtain the updated cumulative dosage. The cumulative dosage at the end of the previous time window is read from the historical dosing record. Updated cumulative dosage .

[0070] Determine whether the updated cumulative dosage exceeds a preset safe dosage threshold. The safe dosage threshold... Set to 100 ml, this value is based on the clinically prescribed upper limit for a single analgesic pump administration (usually 100 ml). When When, it is considered that the threshold has not been exceeded; when When the threshold is exceeded, it is considered to be exceeded.

[0071] When the updated cumulative dosage does not exceed the preset safe dosage threshold, the limited dosing rate is used as the commanded dosing rate at the current moment. That is, the commanded dosing rate. .

[0072] When the updated cumulative total dosage exceeds the preset safe total dosage threshold, the commanded dosing rate is recalculated according to the safe total dosage threshold, and the planned dosage corresponding to the recalculated commanded dosing rate is made to fill the remaining capacity of the safe total dosage threshold.

[0073] When recalculating the commanded dosing rate, the remaining capacity for the safe total dosing threshold is calculated. Let the cumulative total dosing before the start of the current time window be... The safe total dosage threshold is The current time window duration is The length of time that has elapsed is Then the remaining capacity Calculate using the following formula:

[0074]

[0075] Of which: the remaining time is .in, This indicates the length of time, in hours, that has elapsed since the start of the current time window. If the current time is exactly the start time of the time window, then... .

[0076] Recalculated command dosing rate Calculate using the following formula:

[0077]

[0078] in: The unit is milliliters. The unit is hours. The unit is milliliters per hour. The dosing rate is the commanded rate at the current moment.

[0079] In specific implementation, when the commanded drug delivery rate is converted into a stepper motor drive pulse sequence and output to the actuator of the analgesia pump, the drug delivery volume corresponding to each revolution of the stepper motor is obtained. This drug delivery volume is determined by the mechanical parameters of the piston pump driven by the stepper motor, and is set to 0.1 ml per revolution in this system.

[0080] Obtain the number of drive pulses required for one revolution of the stepper motor. The stepper motor has a step angle of 1.8 degrees and requires 200 pulses per revolution, so the number of drive pulses is set to 200 pulses per revolution.

[0081] The required drive pulse frequency per second is calculated based on the commanded dosing rate, the administered volume, and the number of drive pulses. Commanded dosing rate (or The unit is milliliters per hour, which can be converted to milliliters per second: The required drive pulse frequency per second Calculate according to the formula:

[0082]

[0083] Wherein: driving pulse frequency The unit is Hertz.

[0084] The required drive pulse frequency per second is converted into the time interval between adjacent pulses. Time interval The unit is seconds. If ,but Setting it to infinity will not generate pulses.

[0085] A square wave pulse sequence with a fixed pulse width is generated according to the stated time interval. The pulse width is set to 50 microseconds, meaning the high-level duration of each pulse is 50 microseconds and the low-level duration is... Seconds. The starting time of the square wave pulse sequence is the starting time of the current time window, and every second... One pulse is generated per second.

[0086] The square wave pulse sequence is output to the driver input port of the stepper motor. The driver input includes a pulse signal input (STEP pin) and a direction signal input (DIR pin). The direction signal is set to positive, causing the stepper motor to rotate clockwise, pushing the piston in the injection direction. The stepper motor rotates by a corresponding angle according to the number of pulses in the square wave pulse sequence to move the piston of the analgesia pump a corresponding distance. Each pulse causes the stepper motor to rotate 1.8 degrees, completing one revolution every 200 pulses, and each revolution moves the piston by 0.1 ml of drug delivery volume. Specifically, if the total number of pulses output within the current time window is... The volume of medication administered corresponding to the distance the piston moves is milliliters.

[0087] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An adaptive analgesic pump drug delivery control method, characterized in that, include: The patient's real-time physiological parameter sequences and analgesia assessment data were collected as the input dataset; Perform time-frequency transformation on the input dataset to extract pain-related rhythmic fluctuation components and construct a multidimensional time-frequency feature tensor; The feature gradient field is computed on the multidimensional time-frequency feature tensor and manifold learning dimensionality reduction is performed to extract the key manifold structure of the pain event. The key manifold structure is input into a deep reinforcement learning model, and the policy network of the deep reinforcement learning model outputs a baseline drug administration sequence. The baseline drug administration sequence is differentially compared with the real-time physiological parameter sequence to generate a dynamic drug administration adjustment dose; The dynamic dosing adjustment amount is accumulated and fused with historical dosing records to calculate the commanded dosing rate at the current moment; The commanded dosing rate is converted into a stepper motor drive pulse sequence and output to the actuator of the analgesia pump.

2. The adaptive analgesic pump drug delivery control method according to claim 1, characterized in that, The patient's real-time physiological parameter sequences and analgesia assessment data were collected as the input dataset, including: Simultaneously record the patient's heart rate, blood pressure, skin conductance, and blood oxygen saturation signals within a continuous sampling window; The heart rate signal, blood pressure signal, skin conductance signal, and blood oxygen saturation signal are aligned by time and then spliced ​​into a real-time physiological parameter sequence. At the end of each sampling window, obtain the analgesia visual analog scale score triggered by the patient. The analgesia visual simulation score is bound to the real-time physiological parameter sequence of the corresponding sampling window; The bound real-time physiological parameter sequences and analgesia visual simulation scores are organized into a two-dimensional data matrix; The two-dimensional data matrix is ​​used as the input dataset.

3. The adaptive analgesic pump drug delivery control method according to claim 2, characterized in that, Performing a time-frequency transformation on the input dataset to extract pain-related rhythmic fluctuation components and constructing a multidimensional time-frequency feature tensor includes: The heart rate signal is extracted from the two-dimensional data matrix to form a heart rate time-series curve; The skin conductance signal is extracted from the two-dimensional data matrix to form a skin conductance time-series curve; Short-time Fourier transforms are performed on the heart rate time-series curve and the skin conductance time-series curve, respectively, to obtain the heart rate time-spectrum matrix and the skin conductance time-spectrum matrix; The rhythmic fluctuation components with a frequency band between 0.05 Hz and 0.15 Hz are selected from the heart rate spectrum matrix as the heart rate pain rhythm components. In the skin conductance time-spectrum matrix, rhythmic fluctuation components with a frequency band range between 0.1 Hz and 0.5 Hz are selected as conductance pain rhythm components. The heart rate pain rhythm component and the electrical conductance pain rhythm component are stacked in the frequency dimension to construct a three-dimensional time-frequency feature tensor. The first dimension of the three-dimensional time-frequency feature tensor is the time window index, the second dimension is the frequency point index, and the third dimension is the physiological parameter type index.

4. The adaptive analgesic pump drug delivery control method according to claim 3, characterized in that, The process involves computing a feature gradient field on the multidimensional time-frequency feature tensor and performing manifold learning dimensionality reduction to extract the key manifold structure of pain events, including: Calculate the feature difference vector between adjacent time windows along the time window dimension of the three-dimensional time-frequency feature tensor; Combine the feature difference vectors across all time windows into a single feature gradient field; Perform a local linear embedding algorithm on the feature gradient field to project the high-dimensional feature space onto a two-dimensional low-dimensional embedding space; Calculate the local density value of each embedding point in the two-dimensional low-dimensional embedding space; The time window corresponding to the embedding point whose local density value exceeds the preset density threshold is marked as a candidate window for pain events; Extract the corresponding feature vectors of all pain event candidate windows in the three-dimensional time-frequency feature tensor; All extracted feature vectors are organized into a low-dimensional manifold structure, which serves as the key manifold structure for the pain event.

5. The adaptive analgesic pump drug delivery control method according to claim 4, characterized in that, The key manifold structure is input into a deep reinforcement learning model, and the policy network of the deep reinforcement learning model outputs a baseline drug administration sequence, including: Each feature vector in the key manifold structure is sequentially input into the state encoder of the deep reinforcement learning model; The state encoder maps each feature vector to a hidden state vector; The hidden state vector is input into the policy network of the deep reinforcement learning model; The policy network outputs a Gaussian-distributed mean action parameter and variance action parameter. A normalized drug administration action value is obtained by sampling from the Gaussian distribution; Multiply the normalized drug administration action value by the preset maximum drug administration rate to obtain the instantaneous drug administration rate corresponding to the current state; Arrange the instantaneous dosing rates corresponding to all time windows in chronological order to form the baseline dosing action sequence.

6. The adaptive analgesic pump drug delivery control method according to claim 5, characterized in that, The baseline drug administration sequence is differentially compared with the real-time physiological parameter sequence to generate a dynamic drug administration adjustment dose, including: Extract the heart rate, blood pressure, and skin conductance values ​​from three consecutive sampling points prior to the current time window from the real-time physiological parameter sequence. The heart rate, blood pressure, and skin conductance values ​​from the three consecutive sampling points are combined to form a physiological state vector. Extract the baseline dosing rate corresponding to the current time window from the baseline dosing action sequence; Calculate the L2 norm of the vector difference between the stated physiological state vector and the pre-stored ideal physiological state vector; The first adjustment component is obtained by multiplying the second norm by a preset proportional gain coefficient. Calculate the rate difference between the baseline dosing rate and the actual dosing rate in the previous time window; The second adjustment component is obtained by multiplying the rate difference by a preset integral gain coefficient. The first adjustment component and the second adjustment component are added together to obtain the dynamic dosing adjustment amount.

7. The adaptive analgesic pump drug delivery control method according to claim 6, characterized in that, The dynamic dosing adjustment is accumulated and fused with historical dosing records to calculate the commanded dosing rate at the current moment, including: Read the total cumulative dosage at the end of the previous time window from the historical dosing records; Read the real-time dosing rate at the end of the previous time window from the historical dosing records; The dynamic dosing adjustment is superimposed on the real-time dosing rate to obtain the dosing rate to be corrected. The dosing rate to be modified is limited using preset upper and lower limits to obtain the limited dosing rate. Multiply the dose-limited dosing rate by the length of the current time window to obtain the planned dose for the current time window; The planned dosage is added to the total cumulative dosage to obtain the updated total cumulative dosage. Determine whether the updated cumulative dosage exceeds the preset safe dosage threshold; When the updated cumulative total dosage does not exceed the preset safe total dosage threshold, the dosage rate after the limit is used as the commanded dosage rate at the current moment. When the updated cumulative total dosage exceeds the preset safe total dosage threshold, the commanded dosing rate is recalculated according to the safe total dosage threshold, and the planned dosage corresponding to the recalculated commanded dosing rate is made to fill the remaining capacity of the safe total dosage threshold.

8. The adaptive analgesic pump drug delivery control method according to claim 7, characterized in that, The prescribed dosing rate is recalculated according to the safe total dosing threshold, including: Calculate the remaining capacity of the safe total dosage threshold and divide it by the remaining duration of the current time window to obtain the recalculated command dosing rate.

9. The adaptive analgesic pump drug delivery control method according to claim 7, characterized in that, Converting the commanded drug delivery rate into a stepper motor drive pulse sequence and outputting it to the actuator of the analgesia pump includes: Obtain the volume of medication administered per revolution of the stepper motor; Obtain the number of drive pulses required for one revolution of the stepper motor; The required drive pulse frequency per second is calculated based on the commanded drug delivery rate, the drug delivery volume, and the number of drive pulses. The required drive pulse frequency per second is converted into the time interval between adjacent pulses; Generate a square wave pulse sequence with a fixed pulse width according to the time interval; The square wave pulse sequence is output to the driver input port of the stepper motor; The stepper motor rotates by a corresponding angle according to the number of pulses in the square wave pulse sequence to push the piston of the analgesia pump to move a corresponding distance.

10. An adaptive analgesic pump delivery control system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the adaptive analgesia pump drug delivery control method according to any one of claims 1 to 9.

Images